Remove naive rotary forward overriding. (#19263)
This commit is contained in:
@@ -1195,8 +1195,6 @@ class DeepseekV2AttentionMLA(
|
|||||||
scaling_factor = rope_scaling["factor"]
|
scaling_factor = rope_scaling["factor"]
|
||||||
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
|
||||||
self.scaling = self.scaling * mscale * mscale
|
self.scaling = self.scaling * mscale * mscale
|
||||||
else:
|
|
||||||
self.rotary_emb.forward = self.rotary_emb.forward_native
|
|
||||||
else:
|
else:
|
||||||
self.rotary_emb = None
|
self.rotary_emb = None
|
||||||
self.use_deepseek_yarn_rope = rope_scaling is not None
|
self.use_deepseek_yarn_rope = rope_scaling is not None
|
||||||
|
|||||||
Reference in New Issue
Block a user